Step 5a sibling sync: the #2208 satisfiability probe never runs on a SCRIPT-only sync, so a synced sibling script's branch-era-unsatisfiable src/ import reds the gate

Goal

Close the script-only hole in the #2208 import-satisfiability probe. The probe that guards Step 5a's sibling-issue file sync is keyed on synced TEST files only:

for f in "${SIBLING_SYNCED[@]}"; do
  case "$f" in
    tests/test_issue*_*.py)
      if ! (cd "$WT" && timeout --kill-after=15s 180s uv run pytest --collect-only -q "$f" >/dev/null 2>&1); then

(.claude/skills/issue/steps/09-step-5.md, the block around L365-L390.)

The enumeration that FEEDS SIBLING_SYNCED, however, globs scripts as well as tests, and lists only files that DIFFER from origin/main:

done < <(git -C "$WT" -c core.quotePath=false diff --name-only origin/main \
  -- ':(glob)scripts/issue[0-9]*_*.py' ':(glob)scripts/issue[0-9]*_*.sh' \
     ':(glob)tests/test_issue[0-9]*_*.py')

So whenever a sibling issue's SCRIPT has drifted from the branch but its covering TEST is already identical to origin/main, the script is synced and no test enters SIBLING_SYNCED for that issue number — so no probe runs at all. The synced script can then carry a module-level src/ import added to main after this branch's fork point, and the covering (unsynced) test dies at COLLECTION, exactly the #2206/#2208 symptom the probe exists to prevent.

Measured incident (task #2212, 2026-08-20)

Step 10d's pre-push lint gate returned verdict crash — fail-closed, merge blocked — on a branch whose own five deliverable files were clean (the lint legs' baseline and gated normalized outputs were byte-identical at 380 bytes / 3 lines each; lint-new and lint-owndiff both empty).

The crash came from the TG mapped gated leg:

ERROR collecting tests/test_issue2094_figures.py
  tests/test_issue2094_figures.py:27: in <module>
      import issue2094_figures as F
  scripts/issue2094_figures.py:65: in <module>
      from explore_persona_space.analysis.paper_plots import (
E ImportError: cannot import name 'figsize_iclr_full' from
  'explore_persona_space.analysis.paper_plots'
Interrupted: 1 error during collection

Evidence that this is the script-only shape, not #2208's already-fixed test-sync shape:

Fact	Probe
tests/test_issue2094_figures.py is NOT in the branch own-diff (0 paths)	git diff --name-only $MB...HEAD -- tests/test_issue2094_figures.py
It was already identical to origin/main at the merge-base — so it never entered the sync enumeration	git diff --quiet $MB origin/main -- tests/test_issue2094_figures.py → clean
scripts/issue2094_figures.py IS in the branch own-diff — it WAS synced	git diff --name-only $MB...HEAD -- scripts/issue2094_figures.py → 1 path
The sync that introduced it is the sibling-file arm	git log $MB..HEAD -- scripts/issue2094_figures.py → e10ad8b8b4 ... (spec-freshness; sibling-issue files)
main's script carries the module-level import at L66; the merge-base script had zero references	git show origin/main:scripts/issue2094_figures.py vs git show $MB:...
The branch never authored paper_plots.py — it sits at merge-base vintage, which lacks figsize_iclr_full	0 own-commits touching it; grep -c 'def figsize_iclr_full' → main 1, branch 0, merge-base 0

Cost: one full ~45-minute gate run consumed for a crash verdict attributable to sync machinery rather than to the branch payload, plus the diagnosis round. The gate's own single-re-run budget was then spent on the remediation.

Why this is NOT already covered

• #2208 (completed) added the probe for the case where the synced sibling TEST is main-NEW with unsatisfiable src/ imports. Here the test was not synced at all, so its case arm never matched.
• #2412 (running) reports the probe is --collect-only and therefore misses runtime API skew and function-body imports. This incident is the opposite direction: the break IS collect-time, so --collect-only would have caught it — the probe simply was never invoked. Fixing #2412 does not fix this.
• The pair-atomic revert (#1824/#1860) triggers only from a probe FAILURE, so it never fires either.

Step 10d's own prose already names the mechanism as a known hazard — "syncing arbitrary individual test files without their import closure — conftest, tests/ helpers — risks hybrid trees" — but the mitigation was scoped to synced tests.

Candidate fixes (for the planner to adjudicate — not pre-decided)

1. Key the probe on the ISSUE NUMBER, not the synced file's kind. For every distinct <M> appearing in SIBLING_SYNCED, probe every existing tests/test_issue<M>_*.py in the worktree, whether or not that test was itself synced. Smallest change; reuses the existing pair-atomic revert. Cost: one --collect-only per covered issue rather than per synced test.
2. Sync the src/ import closure alongside a synced sibling script. More faithful to what main actually looks like, but widens the sync's blast radius into src/ and needs its own lost-update reasoning — a src/ module is not workflow surface, and blindly checking it out could clobber a genuine branch edit (the family-atomic dirty check would have to extend to it).
3. Revert a synced sibling script whose module-level imports are unsatisfiable, probed directly (python -c 'import <module>' under the worktree venv) rather than via its covering test. Catches the script even when no covering test exists.

Option 1 is the least invasive and closes the measured incident; option 3 generalizes to script/no-test cases. Deciding between them (and whether to do both) is the plan's job.

Acceptance criteria

1. A worktree in which only a sibling SCRIPT has drifted (its covering test identical to origin/main) and whose synced script carries a branch-era-unsatisfiable module-level src/ import either (a) does not commit that sync, or (b) commits it together with a closure that makes collection succeed — verified by a test reproducing the #2212 shape above.
2. The existing #2208 synced-test arm and the pair-atomic revert keep their current behavior (no regression in tests/ pins covering them).
3. The fail-safe direction is preserved: a probe failure or timeout results in status-quo staleness, never a committed hybrid tree.

Provenance

Surfaced during /issue 2212 Step 10d, 2026-08-20, from a real crash verdict on the pre-push workflow-lint gate. Remediated in-branch for #2212 by syncing the import closure (src/explore_persona_space/analysis/paper_plots.py, made byte-identical to origin/main; commit 33f4fb2236) — that unblocked #2212 but does nothing for the next branch to hit this.